Prompting settings

The Prompts tab on an assistant gives you precise control over what Glueprint sends to the underlying coding agent on every wake. It exposes the four prompt layers Glueprint composes (stable context, system prompt overlay, wake message, per-trigger framing) and lets you toggle each piece or replace it outright.

You don’t have to touch any of this for the assistant to work. Defaults are sensible. Open it when you want a different voice, a leaner prompt, or an agent that behaves differently when it wakes from a routine versus from a chat message.

How to open it

1. Open the assistant.
2. Click Settings.
3. Pick the Prompts tab.

Changes save in place and apply on the assistant’s next wake. The Open preview button on the page renders what the assistant would actually receive given the current settings.

The four sections

The tab is organized by the four layers Glueprint composes when an assistant wakes:

1. System prompt mode

Picks whether Glueprint augments, replaces, or yields the harness’s default system prompt.

Three options:

• Append Glueprint to harness (default) — the underlying agent keeps its own system prompt; Glueprint adds an overlay on top.
• Replace the harness system prompt — Glueprint substitutes its own. When selected, a textarea appears for your custom prompt and a checkbox lets you also include Glueprint’s standard overlay below it.
• Harness default only — Glueprint contributes nothing to the system prompt.

The behavior of Replace depends on the underlying agent:

• Claude Code — the custom prompt is appended via the agent’s standard append flag, with full effect.
• Codex — Replace is disabled in the UI with a note that Codex CLI ignores custom system prompts today.
• Gemini CLI — there is no true system-prompt swap, so Glueprint prepends your text as “System instructions:” to the first user message.

2. Stable context (AGENTS.md / CLAUDE.md)

Controls the always-on file Glueprint writes to the assistant’s workspace. Both Claude Code and ACP-style agents read it on every spawn.

Two controls:

• Authority order block (checkbox) — include Glueprint’s explainer that ranks organization rules > per-assistant rules > persona > user notes > memory > recent conversation. On by default.
• Custom preamble (textarea) — free-form text inserted before the Organization Rules section.

3. Glueprint system prompt overlay

A master toggle plus eight section checkboxes. The overlay is what makes a Glueprint assistant behave like a Glueprint assistant (knowing about chat, sub-sessions, memory tools, the workbench, and so on).

• Master toggle — “Include the Glueprint system prompt.” Turn the overlay off entirely if you want a bare agent.
• Section checkboxes — each section can be toggled on or off independently, with a recommendation badge and a one-line hint:
  • Communication channel rule (Required).
  • Safety framing (Recommended).
  • Execution bias (Optional).
  • Sub-session orchestration (Recommended).
  • Memory tools (Recommended).
  • Memory vs. journal taxonomy (Optional).
  • Workbench guidance (Optional).
  • Sub-session tools list (Recommended).
• Custom additions (textarea) — appended after the enabled sections, so you can layer in extra instructions without rewriting the whole overlay.

4. Wake message defaults

A wake message is the per-invocation prompt the assistant gets each time it wakes. It’s composed from blocks; the Wake Message — Defaults section lets you toggle each block on or off:

• Memory (long-term + journal).
• Tasks (open task list).
• Team (if the assistant is on one).
• Due instructions (firing routines).
• Recent conversation (channel history).
• Supervised sessions.
• Dream journals (only relevant for dream-triggered wakes).

Per-trigger overrides

Below the defaults there’s an accordion with six panels, one per trigger type: Chat, Heartbeat, Routine, Task, Mention, Dream. Each panel lets you tailor the wake message specifically for that trigger:

• Section overrides — for every wake-message block, pick Default / On / Off so a heartbeat wake can be leaner than a chat wake, for example.
• Why-you-were-woken framing (textarea) — replaces the built-in copy that explains to the agent why it’s awake.
• Custom additions (textarea) — appended at the bottom of the trigger’s wake message.

A status pill on each panel shows whether the trigger is using Default or lists the active overrides.

Preview

The Open preview button at the top of the tab opens a side sheet that renders:

• The AGENTS.md / CLAUDE.md the assistant will read.
• The system prompt overlay Glueprint will include.
• A sample wake message under the current trigger settings.

Use it to verify a change before you commit to it.

What gets saved where

These settings live on the assistant’s config. Persona and Constitution — the two long-form Markdown documents — are separate; see Persona and constitution.

Related topics

• How Glueprint composes prompts for the architecture, the four layers, and how each harness delivers them differently.
• Persona and constitution for the long-form rules and voice.
• Memory for what’s loaded into the wake briefing.
• Extensions for adding skills, subagents, and MCP servers to the assistant.
• Routines for scheduled triggers that use the per-trigger overrides above.